All-vanadium redox flow battery SOH prediction method based on VMD and integrated neural network

Abstract

The invention discloses an SOH prediction method based on VMD and an integrated neural network. The SOH prediction method comprises the steps that voltage and current under the actual capacity attenuation condition of a battery and voltage and current under the ideal capacity attenuation condition of the battery are obtained through experiments and simulation; voltage and current obtained through experiments and simulation are converted into an SOH change trend for unified processing and analysis; decomposing the SOH change trend into three intrinsic mode functions IMFs and a margin RES (t) based on VMD, and summarizing a fluctuation function F (t) and a main trend function M (t); respectively putting the training sets of the two functions into LSTM and GRU for learning, outputting, adding and verifying a result by using a test set, and forming an integrated neural network algorithm based on VMD; and establishing an SOH prediction model of the VRB based on multiple time scales. The method provided by the invention can effectively face complex and changeable conditions in a long-time operation process, and the precision and stability of the model are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of all-vanadium liquid flow battery management, and more specifically to a method for predicting the health status of a liquid flow battery based on a neural network. Background Art

[0002] With the development of economy and the acceleration of industrialization, the demand for energy is increasing day by day. In order to alleviate a series of environmental problems caused by the development and utilization of fossil energy and ensure sustainable economic development, countries are actively promoting the transformation of energy structure and promoting renewable energy. Vanadium Redox Flow Battery (VRB) is one of the most popular technologies in large-scale energy storage systems. In recent years, it has attracted widespread attention due to its advantages such as low pollution, high safety, long cycle life, and independent power and capacity.

[0003] VRB operation rule is V 2+ / V 3+ and Dissolved in sulfuric acid solution as anode electrolyte and cathode electrolyte respectively. When the battery is charged and discharged, the vanadium ions in the anode and cathode electrolytes undergo redox reactions respectively. Unlike other batteries, the redox products in the VRB electrolyte are non-polluting, do not produce any harmful byproducts, and are system controllable and environmentally friendly. The electrolyte is separated from the battery stack and placed in an independent storage tank, making its power and capacity independent of each other. Although all-vanadium liquid flow batteries show great potential in terms of application prospects, there are still some urgent problems to be solved in their large-scale application stage. Among them, capacity imbalance has become a bottleneck problem restricting the large-scale integration and durable operation of all-vanadium liquid flow batteries. After multiple charge and discharge cycles of all-vanadium liquid flow batteries, the negative electrode electrolyte migrates to the positive electrode, which in turn leads to system capacity imbalance, causing the energy storage system capacity to decay and increase operation and maintenance costs. If not controlled, the capacity will be unbalanced and decayed at the least, or exhausted at the worst. Therefore, most VRB systems require regular maintenance, mainly by mixing or balancing the electrolyte to restore the available capacity. Therefore, it is crucial to predict the health status of VRB, which is not only beneficial to the long-term and stable operation of the VRB system, but also provides a guarantee for the stability and reliability of the VRB energy storage system. Summary of the invention

[0004] In order to fill the research gap in the above-mentioned background technology, the present invention proposes a method for predicting the health status of a flow battery based on a neural network from the perspective of macroscopic mechanism and targeting the capacity regeneration phenomenon.

[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0006] A method for predicting SOH of an all-vanadium liquid flow battery based on VMD and an integrated neural network comprises the following steps:

[0007] Step 1: Obtain the voltage and current of the battery under actual capacity decay and the voltage and current of the battery under ideal capacity decay through experiments and simulations;

[0008] Step 2: Calculate the battery capacity using the ampere-hour integration method using the voltage and current obtained from the experiment and simulation, and analyze the current capacity to the rated capacity SOH to obtain the SOH decay curve;

[0009] Step 3: Decompose the SOH attenuation curve through VMD to obtain three intrinsic mode functions IMFs and a residual RES(t).

[0010] SOH(t)={IMF1(t),…,IMF K (t),RES(t)}

[0011] Among them, IMF i (t)(i=1,2,…K) are the K intrinsic mode function components with appropriate bandwidth after decomposition, RES(t) is the residual component, and K is the number of mode decomposition;

[0012] The fluctuation function F(t) and the main trend function M(t) are summarized through correlation analysis. If the correlation coefficient of IMFi is less than the average correlation coefficient of the three components, it is set to F(t). Otherwise, all IMFi with a coefficient greater than the average coefficient are added to RES(t) and recorded as M(t).

[0013] Step 4: Place the training sets of the fluctuation function F(t) and the main trend function M(t) into LSTM and GRU for learning respectively, and use the test set output to add and verify the results to form an integrated neural network algorithm based on VMD;

[0014] Step 5: Establish a SOH prediction model for VRB based on multiple time scales by changing the neural network internal parameter step size b in the VMD-based integrated neural network algorithm.

[0015] Furthermore, step 1 specifically includes:

[0016] Design and build an all-vanadium liquid flow battery experimental platform, carry out the all-vanadium liquid flow battery capacity decay experiment, and obtain the voltage and current of the battery under the actual capacity decay condition;

[0017] A traditional capacity imbalance model was established using MATLAB R2020b simulation software, and a battery without capacity regeneration was set up for a comparative experiment to obtain the voltage and current under ideal capacity attenuation conditions.

[0018] Furthermore, step 4 specifically includes:

[0019] Repeat the experiment at least 20 times to obtain the reconstruction component F(t) and M(t) data sets, take the first 50% of the data as the training set, and the last 50% of the data as the test set;

[0020] The training set of F(t) is put into the LSTM model for training, and then the F(t) test set is predicted. The prediction of the fluctuation function F(t) is as follows:

[0021] F(t+b)=LSTM(F t-b+1 ,...,F t )

[0022] Where b is the step length, indicating a b-step-ahead prediction;

[0023] The training set of M(t) is put into the GRU model for training, and then the M(t) test set is predicted. The prediction of the main trend function M(t) is as follows:

[0024] M(t+b)=GRU(M t-b+1 ,...,M t )

[0025] Where b is the step length, indicating a b-step advance prediction.

[0026] Further, in step 5, b=1 is the short time scale (STS), b=3 is the medium time scale (MTS), and b=5 is the long time scale (LTS).

[0027] Furthermore, it also includes:

[0028] Step 6: Introduce RMSE as an evaluation indicator to quantify the difference between the predicted value and the true value:

[0029]

[0030] Among them, y i is the actual value, y pre is the predicted value, and n is the length of the time series.

[0031] Compared with the prior art, the advantages of the present invention are:

[0032] (1) The method proposed in the present invention fills the blank part of previous research. For the capacity imbalance problem caused by electrolyte migration in all-vanadium liquid flow batteries, most traditional methods are based on a group of partial differential equations established based on the conservation laws of energy, momentum, mass, charge, etc., and explore the microscopic causes of capacity imbalance and attenuation from the aspects of ion diffusion / migration, water molecule migration, electrolyte / electrolyte imbalance, etc. Although the model has high accuracy, strong extensibility and can explain the evolution of capacity imbalance. However, due to the complexity, high difficulty and high cost of model establishment, it is difficult to apply large-scale laws. Starting from the macroscopic mechanism, the present invention proposes a SOH prediction model based on variational mode decomposition (VMD) and integrated neural network for the capacity regeneration phenomenon. The SOH time series is decomposed by VMD and reconstructed based on correlation to obtain the fluctuation function representing capacity regeneration and the main trend function representing capacity attenuation.

[0033] (2) The present invention uses VMD to decompose and reconstruct the SOH time series signal into the main function M(t) and the fluctuation function F(t), and adopts a multi-scale data-driven method based on the GRU and LSTM integrated model to achieve accurate capacity prediction under different step lengths, which provides an effective means to improve the performance of the VRB system and ensure the long-term reliable and safe operation of VRB.

[0034] (3) Through the SOH prediction method established by the present invention, an integrated prediction model based on GRU and LSTM is established for the main function and the fluctuation function. The method has strong reliability and feasibility. The SOH estimation under different time scales is achieved by changing the size of the step size in the integrated model. The results show that compared with the existing methods, the present invention has high prediction accuracy and strong stability, and the RMSE of the integrated prediction is controlled to be less than 0.5%. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of a method for predicting SOH of an all-vanadium liquid flow battery based on VMD and integrated neural network proposed in the present invention.

[0036] Figure 2 The components and spectra of VMD decomposition of battery 1# at different K values, where (a) and (b) are the components and spectra when K=2, (c) and (d) are the components and spectra when K=3, (e) and (f) are the components and spectra when K=4;

[0037] Figure 3 The reconstructed components F(t), M(t) and integrated predicted SOH result diagram of batteries 1# and 2#. DETAILED DESCRIPTION

[0038] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the embodiments should not be construed as limiting the present invention.

[0039] like Figure 1 As shown, the embodiment of the present invention proposes a method for detecting a fault of an all-vanadium liquid flow battery pump based on a support vector machine, comprising the following steps:

[0040] Step 1: Obtain the voltage and current of the battery under actual capacity decay and the voltage and current of the ideal capacity decay through experiments and simulations. The specific steps are as follows:

[0041] Step 1.1: Design and build a VRB experimental platform to conduct VRB capacity attenuation experiments:

[0042] The VRB experimental platform includes a battery testing system, a host computer and a single battery system, wherein the single battery system includes a single battery, a catheter, a peristaltic pump and a liquid storage tube. The component composition and parameters of the single battery are shown in Table 1 and are named Battery1#.

[0043] Table 1 Single cell composition and parameters

[0044]

[0045] (i) The initial electrolyte volume was 30 mL, and the peristaltic pump flow rate was set to 50 mL / min;

[0046] (ii) Constant current constant voltage (CCCV) charging: 1.8A constant current charging to 1.7V, then constant voltage charging until the current drops to 0.2A, and the battery is left to stand for 1 minute;

[0047] (iii) 1.8A constant current discharge to 0.5V, then constant voltage discharge until the current drops to 0.2A, and then leave for one minute;

[0048] (iv) cycle steps (ii) to (iii) 200 times;

[0049] Step 1.2: Use MATLAB R2020b simulation software to establish a traditional capacity imbalance model and set up a battery without capacity regeneration for comparison. The charge cut-off voltage of the single battery is set to 1.61V, and the discharge cut-off voltage is 1.15V. The charge and discharge are performed at a constant current of 100A, and the current density is 434.8A / m 2 The battery stack is composed of 37 cells, and the volume of the electrolyte in the stack is 0.1m 3The battery operating temperature is controlled at 25°C, and the proton exchange membrane is Nafion115 membrane. The initial capacity of the battery is set to 120Ah and named Battery 2#. Step 2: Calculate the battery capacity after each charge and discharge of the battery according to the ampere-hour integration method, and use SOH (current capacity to rated capacity) for analysis to obtain the SOH decay curve.

[0050] Since the traditional capacity imbalance model is established under the assumption that protons, water molecule migration, electrolyte volume imbalance, etc. are negligible, the SOH decay curve of Battery 2# is relatively smooth. The SOH of Battery 1# continues to increase before the 38th cycle. This is because the battery has not been activated at this time and is currently in the activation period, and the active substances in the electrolyte are not in an active state. Therefore, the capacity of the 38th cycle is used as the rated capacity of Battery 2#, and subsequent calculations only consider data from the 38th cycle onwards instead of the full cycle.

[0051] Step 3: Decompose the SOH decay curve through VMD to obtain the components and main trend function of the SOH time series of Battery 1# when K=3, as shown in Figure 2 shown.

[0052] SOH(t)={IMF1(t),...,IMF K (t),RES(t)} (1)

[0053] Among them, IMF i (t) (i = 1, 2, ... K) are the K intrinsic mode function components with appropriate bandwidth after decomposition; RES (t) is the residual component.

[0054] The working principle of VMD is to first establish a variational constraint model, and then transform the optimized model into a SOH time series through the Hilbert transform, as shown in the following formula:

[0055]

[0056] Where K is the number of modal decompositions, α k (t) is the decomposed intrinsic mode function, ω k It is the IMF i The frequency of , j is the imaginary unit, * is the convolution function, is the first-order time inverse, and δ(t) is the Dirac function.

[0057] Then the Lagrangian operator λ and the quadratic penalty term γ are introduced to solve the variational constraint model, as shown in the following formula:

[0058]

[0059] Among them, <> is the inner product algorithm, u k ,ω k , the parameters of λ are updated by the Lagrangian function. According to Parseval's theorem, by updating the initialization parameters λ 1 (ω), and the iterative parameters are as shown in the following formula.

[0060]

[0061] Repeat the above process until the iteration stop condition is met.

[0062]

[0063] Among them, ε is the accuracy determined at the end of the iteration.

[0064] Through correlation analysis, the IMF components are summarized and reconstructed into the main trend function M(t) and the fluctuation data signal F(t). i If the correlation coefficient of is less than the average correlation coefficient of the three components, it is set to F(t). Otherwise, all IMFs with coefficients greater than the average coefficient are set to i It is accumulated with RES(t) and recorded as M(t).

[0065] Calculating IMF i The Pearson correlation coefficient between the original curve and the average correlation coefficient is obtained to determine whether it is a fluctuation function. Table 2 gives the Pearson correlation coefficients of each component of battery 1# and the average correlation coefficient.

[0066] Table 2 Correlation coefficients of eigenmode functions

[0067] Modal Components Pearson correlation coefficient <![CDATA[IMF1]]> 0.0501 <![CDATA[IMF2]]> 0.9976 <![CDATA[IMF3]]> 0.9695 Average 0.6724

[0068] It is found that the correlation coefficient of IMF1 is lower than the average coefficient, so IMF1 is considered to represent capacity regeneration and is named the fluctuation function F(t). All the remaining components are added together to form a component whose change trend is very close to the original function. It is considered to be the SOH decay curve that does not include capacity regeneration and is named the main trend function M(t).

[0069] M(t)=IMF2(t)+IMF3(t)+RES(t) (8)

[0070] The SOH after comprehensive prediction by the integrated model is obtained as shown below.

[0071] SOH(t)=F(t)+M(t) (9)

[0072] Step 4: Place the training sets of the fluctuation function F(t) and the main trend function M(t) into LSTM and GRU for learning respectively, and use the test set output to add and verify the results to form an integrated neural network algorithm based on VMD. Step 4 specifically includes:

[0073] The experiment was repeated 20 times to obtain the reconstruction component F(t) and M(t) data sets. The first 50% of the data was taken as the training set, and the last 50% of the data was taken as the test set. The training sets of the reconstruction components F(t) and M(t) were respectively placed in LSTM and GRU for training, and then the test sets were predicted. Finally, the test sets of the two models were integrated to output the SOH prediction value, which was compared with the true SOH value.

[0074] Through the collaboration of LSTM multiple gates, it is possible to effectively deal with data with large fluctuations. The prediction of the fluctuation function F(t) is as follows:

[0075] F(t+b)=LSTM(F t-b+1 ,...,F t ) (10)

[0076] Based on the GRU algorithm with memory capability, the prediction of the main trend function M(t) is as follows:

[0077] M(t+b)=GRU(M t-b+1 ,...,M t ) (11)

[0078] Where b is the step length, indicating a b-step advance prediction.

[0079] Step 5: Establish a SOH prediction model for VRB based on multiple time scales by changing the neural network internal parameter step size b in the VMD-based integrated neural network algorithm.

[0080] Among them, b = 1 is short time scale (STS), b = 3 is medium time scale (MTS), and b = 5 is long time scale (LTS). By changing the parameter step size b in the neural network, a SOH prediction model based on VRB with multiple time scales is established.

[0081] Step 6: Introduce RMSE as an evaluation indicator to quantify the difference between the predicted value and the true value.

[0082]

[0083] Among them, y i is the actual value, y pre is the predicted value, and n is the length of the time series.

[0084] Figure 3The reconstruction component prediction and integrated prediction results of the two batteries are shown. It is found that both the reconstruction component and the integrated component can be accurately predicted, and the integrated prediction RMSE of the two batteries can be controlled within 0.5%.

[0085] Those skilled in the art should understand that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting SOH of all-vanadium liquid flow battery based on VMD and integrated neural network, characterized in that: The steps include: Step 1: Obtain the voltage and current of the battery under actual capacity decay and the voltage and current of the battery under ideal capacity decay through experiments and simulations; Step 2: Calculate the battery capacity using the ampere-hour integration method using the voltage and current obtained from the experiment and simulation, and analyze the current capacity to the rated capacity SOH to obtain the SOH decay curve; Step 3: Decompose the SOH attenuation curve through VMD to obtain three intrinsic mode functions IMFs and a residual RES(t). SOH(t)={IMF1(t),…,IMF K (t),RES(t)} Among them, IMF i (t)(i=1,2,…K) are the K intrinsic mode function components with appropriate bandwidth after decomposition, RES(t) is the residual component, and K is the number of mode decomposition; The fluctuation function F(t) and the main trend function M(t) are summarized through correlation analysis. If the correlation coefficient of IMFi is less than the average correlation coefficient of the three components, it is set to F(t). Otherwise, all IMFi with a coefficient greater than the average coefficient are added to RES(t) and recorded as M(t). Step 4: Place the training sets of the fluctuation function F(t) and the main trend function M(t) into LSTM and GRU for learning respectively, and use the test set output to add and verify the results to form an integrated neural network algorithm based on VMD; Step 5: Establish a SOH prediction model for VRB based on multiple time scales by changing the neural network internal parameter step size b in the VMD-based integrated neural network algorithm.

2. The method for predicting SOH of all-vanadium liquid flow battery based on VMD and integrated neural network according to claim 1, characterized in that: Step 1 specifically includes: Design and build an all-vanadium liquid flow battery experimental platform, carry out the all-vanadium liquid flow battery capacity decay experiment, and obtain the voltage and current of the battery under the actual capacity decay condition; A traditional capacity imbalance model was established using MATLAB R2020b simulation software, and a battery without capacity regeneration was set up for a comparative experiment to obtain the voltage and current under ideal capacity attenuation conditions.

3. The method for predicting SOH of all-vanadium liquid flow battery based on VMD and integrated neural network according to claim 1, characterized in that: Step 4 specifically includes: Repeat the experiment at least 20 times to obtain the reconstructed component F(t) and M(t) data sets, take the first 50% of the data as the training set, and the last 50% of the data as the test set; The training set of F(t) is put into the LSTM model for training, and then the F(t) test set is predicted. The prediction of the fluctuation function F(t) is as follows: F(t+b)=LSTM(F t-b+1 ,...,F t ) Where b is the step length, indicating a b-step-ahead prediction; The training set of M(t) is put into the GRU model for training, and then the M(t) test set is predicted. The prediction of the main trend function M(t) is as follows: M(t+b)=GRU(M t-b+1 ,...,M t ) Where b is the step length, indicating a b-step advance prediction.

4. The method for predicting SOH of all-vanadium liquid flow battery based on VMD and integrated neural network according to claim 1, characterized in that: In step 5, b=1 is the short time scale (STS), b=3 is the medium time scale (MTS), and b=5 is the long time scale (LTS).

5. The method for predicting SOH of all-vanadium liquid flow battery based on VMD and integrated neural network according to claim 1, characterized in that: Also includes: Step 6: Introduce RMSE as an evaluation indicator to quantify the difference between the predicted value and the true value: Among them, y i is the actual value, y pre is the predicted value, and n is the length of the time series.

Images